Learning-based image compression is described. In one implementation, an encoder possessing a first set of learned visual knowledge primitives excludes visual information from an image prior to compression. A decoder possessing an independently learned set of visual knowledge primitives synthesizes the excluded visual information into the image after decompression. The encoder and decoder are decoupled with respect to the information excluded at the encoder and the information synthesized at the decoder. This results in superior data compression since the information excluded at the encoder is dropped completely and not transferred to the decoder. Primitive visual elements synthesized at the decoder may be different than primitive visual elements dropped at the encoder, but the resulting reconstituted image is perceptually equivalent to the original image.